Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""H3's host-placed prequant pickle is mapped on Linux; ``UNSLOTH_DIFFUSION_PREQUANT_MMAP=0`` restores the full read."""
import ast
import inspect
from pathlib import Path
import torch
BACKEND = Path(__file__).resolve().parents[1]
def test_host_destination_maps_only_when_a_host_copy_follows_on_linux(monkeypatch):
from core.inference import diffusion_prequant as pq
monkeypatch.delenv("UNSLOTH_DIFFUSION_PREQUANT_MMAP", raising = False)
monkeypatch.setattr("sys.platform", "linux")
assert not pq.prequant_mmap_enabled("cpu")
assert pq.prequant_mmap_enabled("cpu", host_copy_follows = True)
assert pq.prequant_mmap_enabled("cuda", host_copy_follows = False)
monkeypatch.setattr("sys.platform", "win32")
assert not pq.prequant_mmap_enabled("cpu", host_copy_follows = True)
monkeypatch.setattr("sys.platform", "linux")
monkeypatch.setenv("UNSLOTH_DIFFUSION_PREQUANT_MMAP", "0")
assert not pq.prequant_mmap_enabled("cpu", host_copy_follows = True)
def test_mapped_host_read_matches_the_full_read(tmp_path, monkeypatch):
from core.inference import diffusion_prequant as pq
monkeypatch.delenv("UNSLOTH_DIFFUSION_PREQUANT_MMAP", raising = False)
monkeypatch.setattr("sys.platform", "linux")
ckpt = {
"format": pq.PREQUANT_FORMAT,
"state_dict": {"w": torch.arange(64, dtype = torch.int8).reshape(8, 8), "s": torch.rand(8)},
"metadata": {"scheme": "int8"},
}
path = tmp_path / "x.pt"
torch.save(ckpt, path)
seen = []
real = pq._load_prequant_checkpoint
def spy(p, **kwargs):
seen.append(kwargs.get("mmap"))
return real(p, **kwargs)
monkeypatch.setattr(pq, "_load_prequant_checkpoint", spy)
mapped = pq._read_prequant_for(str(path), "cpu", host_copy_follows = True)
full = pq._read_prequant_for(str(path), "cpu")
assert seen == [True, None]
for key, value in ckpt["state_dict"].items():
assert torch.equal(mapped["state_dict"][key], value)
assert torch.equal(full["state_dict"][key], value)
def test_loader_forwards_the_hint():
from core.inference import diffusion_prequant as pq
params = inspect.signature(pq.load_prequantized_transformer).parameters
assert params["host_copy_follows"].default is False
src = inspect.getsource(pq.load_prequantized_transformer)
assert "host_copy_follows = host_copy_follows" in src
def _h3_prequant_call() -> ast.Call:
tree = ast.parse((BACKEND / "core/inference/video.py").read_text(encoding = "utf-8"))
for node in ast.walk(tree):
if isinstance(node, ast.FunctionDef) and node.name == "_load_h3_modular_pipeline":
for call in ast.walk(node):
if (
isinstance(call, ast.Call)
and isinstance(call.func, ast.Name)
and call.func.id == "load_prequantized_transformer"
):
return call
raise AssertionError("no load_prequantized_transformer call in _load_h3_modular_pipeline")
def test_h3_load_passes_the_hint_from_its_render_device():
call = _h3_prequant_call()
kw = {k.arg: k.value for k in call.keywords}
assert ast.unparse(kw["device"]) == "'cpu'"
assert ast.unparse(kw["host_copy_follows"]) == "_h3_host_copy_follows(device)"
def test_h3_host_copy_follows_only_on_a_discrete_gpu(monkeypatch):
from core.inference import host_pinning
from core.inference import video as V
monkeypatch.setattr(host_pinning, "device_is_integrated", lambda device = None: False)
assert V._h3_host_copy_follows("cuda")
assert V._h3_host_copy_follows("cuda:1")
assert not V._h3_host_copy_follows("cpu")
assert not V._h3_host_copy_follows(None)
monkeypatch.setattr(host_pinning, "device_is_integrated", lambda device = None: True)
assert not V._h3_host_copy_follows("cuda")